Abstract: As a language spoken by tens of millions of people,Thai is widely used.In the late 1990s,some scholars carried out research on Thai speech synthesis.In recent years,end-to-end speech synthesis systems based on deep neural networks and trained with large-scale high-quality “text-audio” data have been able to synthesize high-quality speech.At present,Chinese,English and other common languages have massive speech synthesis databases.However,the “text-audio” database available for Thai as a non-common language is often small in scale.Under the condition of low resources,this paper aims to improve the quality of Thai speech synthesis,selects the end-to-end speech synthesis model Tacotorn2 as the baseline model,studies the alternate training method and pre-training method,and studies the effect of different text embedding methods on the effect of Thai speech synthesis.Then,the speech synthesized by the six models designed in this paper is evaluated from the attention alignment map and the MOS score.Experimental results show that the system using the method of “vowel consonant embedding+pre-training+alternate training” has the best speech synthesis quality,and the MOS score of the synthesized speech can reach 3.95,which is significantly better than the baseline system’s 1.71.

Key words: Speech synthesis, Thai, Low resource, Alternate training, Pre-training